\(CM_R\), Ridgway

Some of the most ambitious attempts at mapping colors to their names, or naming colors, came from the natural sciences. American ornithologist Robert Ridgway (1850-1929), for example, authored two influential works of color naming systems: A Nomenclature of Colors for Naturalists in 1886, and Color Standards and Color Nomenclature in 1912. In the preface of the earlier work, Ridgway names as his problem that “the author has in collection considerably over three hundred water-colors, each bearing a different name” (Ridgway, A Nomenclature of Colors for Naturalists X). This volume contains over a thousand colors, but the color names are less metaphorical (lemon-yellow), and more descriptive (bright yellow).

\(CM_S\), Saccardo

Another influential work of color naming is the ambitious and polyglot volume from Italian botanist Pier Andrea Saccardo bearing the formidable Latin title, Chromotaxia Seu Nomenclator Colorum, Polyglottus, Additis Specimibus Coloratis ad Usam Botanicorum et Zoologorum (1894). Although containing only fifty colors, it features an index of several hundred “synonyms” for these colors in Latin, Italian, French, English, and German. While some of these are recognizable to modern readers, others seem strangely specific, such as Murinus (mousey) or Fuligineus (“sooty”). Saccardo provides two supplementary colors: achrous, or colorless, glassy; and sordidus, or “sordid,” “dirty,” which he describes as a modifier rather than a color. “non est color definitus sed indicat inquinamentum aliorum colorum. Exempla: sordide albus, luride ruber” (Saccardo 16).

\(CM_M\), Maerz and Paul

Maerz and Paul’s 1930 A Dictionary of Color provides the largest number of colors, and was itself meticulously compiled from a number of prior manuals. This volume contributes over three thousand colors.

\(CM_P\) Pantone

The Pantone set, one of the most common among designers and artists today, contains over two thousand colors, but they have names which are much more mercantile than others. Thus, these are biased towards food-related words, flower-related words, or anything else that would seem like a pleasant marketing term.

\(CM_X\), XKCD

The antidote to the Pantone set is one from Russell Monroe, an American author, former NASA engineer, and cartoonist best known for his webcomic XKCD. Monroe surveyed his wide readership, asking them to name colors they were shown at random on his website. He also took demographic data from them, logged their locations via their computers’ addresses, and asked them whether they were colorblind, or used a cathode ray tube monitor. The survey results, which represent the five million color mappings from 220,500 users, show a consensus for many color names, as shown in table 1.I’ve represented these hex values in RGB space using a script which displays them in the browser, overlaid on the hex value itself, to account for color differences in monitors, pages, and other media.

This mapping presents a useful counterpoint to commercial mappings such as that of Pantone, or to more systematic mappings like Ridgway’s. In the sample presented in [fig:xkcdBlocks], we see a mix of naming metaphors. The usual food metaphors (melon, mocha, coffee) appear next to animal metaphors (camel, canary yellow) and creative compounds indicating a small amount of one color mixed into another (purplish, bluey, purply, preyish). The informality of the “-ish” suffix suggests extemporaneous description, as if colors are mixing in the imaginations of these survey respondants, in the absence of a ready-made metaphor. For comparison, greyish pink in this color map is blush in Pantone, and darkish green translates to online lime. And of course, one would expect that sickly green would not be an easily marketable name for a commodity, especially if it were food, so in Pantone the color is lime green. If an exact match for a hex value does not exist in a color map, I find the closest color to it using \(\Delta E^{*}_{ab}-76\). This is described in more detail in Categorization.

Summary and Comparison

To compare the tendencies, or biases, of these color maps, and to better know how to balance them, I calculate the average of their 300-dimensional GloVe vectors (Stanford’s Global vectors for word representation, trained on English-language websites), and derived the cosine similarity to the vectors of a number of seed words:

\[similarity(\vec{A}, \vec{B}) = \frac{ \vec{A} \cdot \vec{B} }{\|\vec{A}\| \times \| \vec{B} \|}\]

Or, the dot product of the two vectors, normalized by the product of their two \(L_2\) (Euclidean) norms.

Fig. 3 shows a series of word vectors chosen to illustrate the vector similarity with the average vectors of each color map. \(CM_P\), the Pantone map, has a higher similarity to positive, marketable-sounding words, and words evoking leisure, whereas \(CM_X\) has a higher similarity for snot, a decidedly unmarketable word, discussed later, and Jaffer’s aggregation of \(CM_S, CM_R, and CM_M\) shows a slighly higher similarity for blood, also not a decidedly marketable term.

Word Proximity-based, \(M_P\)

One of the simplest methods of color inference is to calculate the syntactic distance from a known color word to a target word. Given a large enough corpus, it is quite likely that, for example, green will appear within several words of grass, and so by measuring the distances from green to grass, and noticing that these distances are much shorter than for the pair red and grass, we might infer that grass is green: that is, the literary imagination of grass has the color category green.

Another example might be inferring the color of a gull. Anyone who has visited the north Atlantic shore knows that gulls tend to be white and gray. Of the 170 times that the word gull appears in \(C_{PG}\), we see white appear within about ten words of it nineteen times. The lemma grey appears six times. Red, however, appears not once. Of course, both green and yellow appear twice, although not with the same relations in the dependency graph. Given this collocation data, we can write a model that guesses that gull is mostly white, a little gray, and with hints of green and yellow.

However, syntactic proximity is preferable to raw proximity itself, and so I developed an algorithm to score relations between two neighboring words, which uses both linear word distances and syntactic distances. I calculate syntactic distances by traversing the dependency trees of their containing sentences. By way of illustration, take these lines from Arthur Conan Doyle’s 1908 novel Sir Nigel:

Next morning they found themselves in a dangerous rock studded sea with a small island upon their starboard quarter. It was girdled with high granite cliffs of a reddish hue, and slopes of bright green grassland lay above them. (Doyle 250)

The syntax dependency graph of the clause, “slopes of bright green grassland lay above them” is parsed as shown in fig. 4

Here, bright and green are descendants—syntactic dependents—of grassland. This might even more accurately be parsed with bright and green together as one semantic unit.

This model infers color associations \(W_C\) from target words, \(W_T\), by traversing the syntax tree, and calculating weights accordingly.

However, modifiers are not always direct descendants of their modified words, since they might cross sentences. (Imagine a passage that were to read: “The slopes of grassland. How bright green they were!”) So to account for these types, I also compute weights based on the raw distances of these words from each other. The full algorithm is this:

1. It begins by identifying a color word, \(W_C\) from color map \(CM_X\) in the target text.
2. It then parses the containing sentence, and determines its syntactic dependencies.
3. Starting from \(W_C\), it navigates through parent words and parent noun chunks \(W_T\) to the root of the sentence.
4. If \(W_T\) is a noun or adjective, it is assigned a score: 2 if it is a direct parent of \(W_C\), or 1 if it is a grandparent of \(W_C\), at two steps’ removal in the syntactic tree.
5. All other words nouns and adjectives are now candidate $W_T$s, and are assigned a score: \(1/i\) where \(i\) is the distance, in number of tokens, from \(W_C\). Thus, it gets a score of 1 if it is a directly adjacent word, or 0.5 if it is two tokens away.
6. These scores are then averaged for each token that shares the same lemma.

The resulting data structure looks like fig. 5, for grass:

There are plenty of colors, like blue, which are uncommon for grass (except for the Appalachain style of music). Yet these other colors represent only a small proportion of that for green. I blend these colors together, by finding their hex values in \(CM_X\), and then averaging their RGB values, after weighting them, so given RGB values \(R_i ... R_j\), \(G_i\), and \(B_i\), and weights \(W_i\), averaging them proportionally:

\[blended RGB = \frac{ W \sum_{R_i}^{R_j} }{ \sum_{i}^{j} W } , \frac{ W \sum_{G_i}^{G_j} }{ \sum_{i}^{j} W} , \frac{ W \sum_{B_i}^{B_j} }{\sum_{i}^{j} W}\]

Or, for all colors, weighting and summing each component of RGB space, and then dividing by the total sum of all weights. For grass above, we get #6b9c56, a pleasantly grassy color.

This model works reasonably well—that is, meets our modern English-speaking expectations of the archetypal colors of many nouns. But there are quite a few notable differences. One is the principle of color description exceptionalism I outline above: color descriptions are anomalies. Although sheep are typically white, and black sheep comparatively rare, the probability of encountering the phrase black sheep in British fiction is about five times greater than that of encountering white sheep, according to bigram data from \(C_{NG}\).

Here’s a portion of the model’s inferences for sheep:

The blended color from \(CM_X\) is #5B5441, a dark greenish, and hardly the color one would expect of a sheep. Part of this is because, as you can see from the inferences above, \(CM_X\) contains many metaphoric colors like grass, heather, and stone. Although these are almost certainly not used as color words in their original contexts, this model counts them as colors. This is perhaps not a bug, however, but a feature of the program: by picking up on elements like stone and grass, we might fail to imagine the sheep itself, but we succeed in imagining the hue of its context, and in a sense, this is more information than we would get from simply imagining a sheep. And after all, when we ourselves imagine a sheep, we might very likely imagine it in its pastoral context, among grass and stones, rather than floating in a colorless void, away from everything else.

But the converse of that phenomenon is also present. We might not expect to see white and gray occur so often close to gull, because these are obvious descriptors, yet we do. When a writer wants to underscore the visual properties of an object that already has well-known color properties, this contributes to the impressionism of the piece: it allows the reader to see through the writer’s eyes.

This phenomenon is brilliantly at play in another Katherine Mansfield story, “Bliss,” written two years earlier in 1918. It is even more colorful than “The Garden Party,” and its colors much more variegated. This reflects the mood of its protagonist, Bertha, who is so happy that she is almost manic. Her attentions are everywhere, and glittering, reflecting off of everything.

There is a very particular quality of light in this story: it is not a categorical color, but a phenomenological one, describing not what Bertha knows the color to be, but how it seems. Bertha sees and feels bright sparks. The narrator tells us that “in her bosom there was still that bright glowing place—that shower of little sparks coming from it” (Mansfield 145). This moment is mirrored in a description of a fruit bowl which Mary at that moment brings in, and she perceives as “a glass bowl, and a blue dish, very lovely, with a strange sheen on it as though it had been dipped in milk.” Here, Mansfield contrasts the color category of the dish—it is blue—with its percepual reality: it appears white.

Bertha sees the fruit in this bowl also with an acute sense of newness: not quite estrangement, in Shklovsky’s formulation, but a newly intimate familiarization. “There were tangerines and apples stained with strawberry pink. Some yellow pears, smooth as silk; some white grapes covered with a silver bloom and a bug cluster of purple ones. These last she had bought to tone in with the new dining-room carpet” (146).

“Apples stained with strawberry pink,” is an uncommon metaphor for apples, which \(M_P\) models as:

And for which the blended color is #AC883B. In other words, apples are usually golden, green, red, gold, or yellow, but rarely strawberry pink. (There is an apple cultivar called the “pink lady,” but it wasn’t cultivated until 1976.) As with milk, this is a food metaphor, one which uses these colors, as the foods themselves seem to do, as a marker for something delicious to come.

Grapes, too, typically belong to the categories red or white, much like their wine. Yet white grapes, and white wine, are not white at all, but a pale green. Red grapes and red wine are not red, either, but are usually a deep purple.

Mansfield would have been aware of the categories of grapes, but has Bertha call them “purple” to show that she is attune to the phenomenon of their hue, and further underscores this by showing the chromatic harmony Bertha expected, and sees, between the grapes and the carpet.

So by modeling imagination in this way, we’re able to take advantage of the ways writers show us the perceptual experience of words, rather than just their categories. But this model is only one component of a larger system.

Dictionary-based inference, \(M_D\)

I also mine data useful for color inference from more straightforward definitions, like those found in dictionaries and encyclopedias. Project Gutenberg provides a copy of Chambers’s Twentieth Century Dictionary of the English Language, published in London in 1908. If I did not already have color mappings for grass (\(CM_X\): grassy green, #419c03; grass, #5cac2d; grass green, #3f9b0b), we would find this entry for grassy: “covered with or resembling grass, green.” Since words and their definitions are regularly formatted in this dictionary, it becomes possible to parse the dictionary entry into word/definition pair, and load them into a lookup table. From there, I construct an graph, where nodes are dictionary words and their colors, and edges accrue weight each time a color word from the color maps appears in the definition of one of the defined words. Thus, if we see grassy appear on the left of the page, and green in its definition, we give it a score of one. If green were to appear twice in its definition, we’d give it a score of two, and so forth. I repeat this process an encyclopedia: The Nuttall Encyclopædia, first published in London in 1900, and still in print by 1966.

Other dictionary and encyclopedia-like works are available in the usual plain-text sources, but few meet the criteria of (a) British, (b) out-of-copyright, (c) in plain-text, and (d) in an easily parsable format.

This model does not seem to perform as well as \(M_P\), however. Here is an entry for grass, for example.

The model triangulates between the color word and entry/definition pairs, but only finds a few single instances each. Possibly as a happy coincidence, however, the resulting aggregate here is #89bc6e, a very grassy color.

When merging these models, I weight \(M_D\) the lowest, since with a lack of data (dictionaries), and an inconsistent level of descriptive detail, I don’t trust it as well as less declarative models.

Image-based: \(M_I\)

A third solution mitigates some of the problems described above, such as the “no white sheep” issue, by escaping comparison between colors and objects, and looking to images as a source of color information.

Given a database of images that are correctly labeled, it should be possible to extract color values from those images, and build a pipeline that infers a color given a word. Lucily, two somewhat newly-created web services provide such a database: Unsplash and Pexels are stores of open-licensed stock photos and illustrations, which provide open APIs [Application Programming Interfaces] which allow users to retrieve images based on a given keyword.

However, one problem that arises here is that an image of a given object, such as a sheep, rarely contains only sheep. There is almost always a background to the image which is of a different color. I try to control for this difficulty by disregarding the most frequent color of these images, and only working with the second and lesser frequent colors. From there, I average the resulting colors of all the images, to retreive an imagined (inferred) hue for the lemma.

The resulting algorithm may be summarized as follows:

1. Scan through the text, looking for nouns, adjectives, or similar words (words with potential visual content).
2. For each matching word, find its lemma.
3. Query the Pexels API, giving the lemma as a search term, and ask for ten image addresses. Download them.
4. For each downloaded image, find the second to Nth most frequent color.
5. Average all these colors proportionally, using the algorithm described above.
6. Return a pairing of a word with a corresponding RGB hex value.

Here is a sample of the model’s inferences for three fairly concrete words, wheat, butterfly, and tennis, and with three more abstract words, deposit, pure, and travelling. I’ve included a few images used in the creation of each color, to allow for some model introspection.

wheat

butterfly

tennis

deposit

pure

travelling

Wheat the model predicts reasonably well: the resulting color is a golden brown. Butterfly is a little more uncertain, in part because there are many species of butterfly—tens of thousands—and there is an incredible diversity of color patterns between them. Tennis one might have expected to be more greenish, since most tennis courts are green. The averaged color here reflects the inclusion of the image where the court is red, and the image of the racket alone. Deposit is curious: probably because it would be a rare word to use to tag a photo, the first four images are likely from the same photographer and shoot. The last is an image of bars of gold, however. This is interesting, since it is not a stack of paper money, but of an archetypal idea of money—the gold standard—which is no longer in widespread use. The inferred color for pure, interestingly, is anything but—it is a dirty grayish. Even though most of these images are water-related, one is of honey. This seems to show one of the pitfalls of computationally imagining an abstract concept. Travelling, on the other hand, is a very distinct blue, owing to the color of the beach scenes that have become emblematic of travel today. Coincidentally, however, blue seas would have been a common feature of international travel in the early twentieth century.

A broader view of this model, shown below, gives one a sense of a few trends. First, almost all of these colors are very desaturated: they appear as if they began as pure pigments, and were mixed with a titanium white. This is owing to the way each of these is by nature an average of several colors: they are mixed. At a glance, the legal-sounding words power and possession appear darkest. The two words dealing with negative emotion, rudeness and objection, appear redder than the others. And words that evoke comfort, reliance, liberty, and amusement, appear bluish. Neighbour and neighbourhood, meanwhile, appear green, probably owing to stock images that involve green lawns.

When projected in HSL space, certain patterns emerge, as well, as shown in fig. 6.

For the problems cited above, I weigh this model higher than \(M_D\) but still lower than \(M_P\). In the absence of an evaluation metric—and I doubt one is possible—I feel like our intuition as readers, and imaginers, is enough to evaluate and weigh these methods of imagining.

Comparison

I combine these three models according the weights given in table 3, and produce a master mapping I’m calling a deep imaginer. This I then integrate with the shallow color maps described above, cascading all the models together.

The Rose Problem

Early iterations of this engine found a staggering incidence of the color word rose in the literature of this period. It didn’t take much introspection to determine that this was not exclusively the color rose, but either a woman’s given name, Rose (my parser is case-insensitive), an equivalent surname, or the past tense of the verb rise. One could remove all tokens where the first letter is capitalized, but then that would also eliminate the color rose which appears at the beginning of a sentence. One could run a part-of-speech tagger over the text, and throw out all the verbs, but this only solves half the problem.

A method exists which, if properly extended, mitigates this problem. Named Entity Recognition, or NER, is a sub-field of natural language processing which computes the probability that a string of words is a named entity. Usually, these are people, places, organizations, languages, ethnicities, and so on, and detecting them has been useful in commercial applications of text analysis, where a corporation might be interested in finding all instances of Apple, the computer company, but ignore all instances of apple, the fruit.

NER has been practiced, in one form or another, since at least Lisa Rau’s 1991 paper, Extracting Company Names from Text (Rau). But whereas early techniques like Rau’s were heuristic, modern methods use computational neural networks to achieve this end. Among the most accurate NER engines now is Explosion AI’s SpaCy library, which uses residual convolutional neural networks along with specialized word embeddings to achieve reasonably accurate predictions of entities like personal names, organization names, as well as disambiguate numeric tokens into quantities, cardinal numbers, and so on (Honnibal).

NER becomes useful to my study in two ways: first, it allows me to discard those entities like “Rose,” (a given name), as well as “Mrs. Brown,” and “Mr. Green.” (Although there is a case to be made that the writer’s choices of these names, where they are fictional, are as deliberate as choice as a visual descriptor.) Second, since this general-purpose NER doesn’t detect color expressions, descriptions, or any literary features, I train it to do so.

I train two NER models, which I’ll use to detect text with visual properties: a shallow model, \(NER_S\), and a deep model \(NER_D\). The seed data set for the more restricted model, \(NER_S\), is trained on raw color entities from \(CM_X\), while that for \(NER_D\) is trained on the set of more general visual expressions spans generated from the deep imaginer.

The Beret Test

These are not the only training data sets I feed to the model, however. I use SpaCy’s Prodigy tool to dynamically correct the model’s most uncertain guesses, across a corpus of edge-cases. In training \(NER_S\), I aim for a more restricted definition of a color: one which is parsable as a color without much context. For example, in the famous hook from the Prince song, “Raspberry Beret,” we understand that in the words “she wore a raspberry beret,” “raspberry” is the color of her beret, which is made of cloth, and not of one or more raspberries. But if we replace raspberry with a more uncommon color, say, electric, or cyberspace (color names from \(CM_P\)), it is not still clear that the word is a color description, and does not describe some other attribute.

The process of training this model was instructive. \(NER_S\) found instances of the word salmon often, and since the word appears in nearly every color map (e.g., \(CM_R\): #D9A6A9; \(CM_X\): #FF796C), it identifies any use of salmon, regardless of whether it refers to a color. It is important to note here that salmon, the color, refers to the color of salmon meat, rather than the color of the fish itself. In other words, the color refers to the inside of the fish, rather than the outside. Disambiguating between these two senses of the word is a non-trivial task for this model. Nonetheless, given a large number of training examples, the model is able to perform better than chance.

What we are left with, after training, are probabilistic models capable of identifying explicit color expressions, in the case of \(NER_S\), and anything that may be imagined, in the case of \(NER_D\). So, to employ Mansfield again, \(NER_S\) finds instances of green baize, and assigns it an RGB value, whereas \(NER_D\) finds instances of tennis lawn, and also assigns it a green value (following color inference, described below).

But first, we would need a way of categorizing the color values generated by these models. Much in the same way that a text analysis project needs a lemmatization system to group together words like sky and skies, go and went, we need a way to bring together variations, shades, and other common properties of colors.

Color Spaces and Color Difference

We now move from categorical description of color to quantitative—i.e., from that is blue to that is 80% blue. Now that we’ve mapped color words and other color-containing text to the model’s guesses of their mappings, we need a way to organize them, in relation to one another, and in relation to our perception of them. This is problematic, since we are dealing with two ontological domains: a linguistic domain, and an analogous psycho-physical. Furthermore, there are a multitude of ways to quantify color properties, and to organize colors by those properties.

Relations between colors—color difference—is a long-standing problem. Colors are typically categorized in relation to one another by embedding them in a color space: a vector space in which each color is a point in its coordinate system. These spaces are very precisely described in technical literature (Fairchild, for instance). The biggest problem they attempt to solve is that hues themselves, due to the anatomy of the eye, do not have linear relationships—this is why three-dimensional projections of these color spaces are often conical, or even asymmetrical. Furthermore, each color space must account for ocular physiology across individuals; differences in ambient reflectors, illuminants, and other lighting conditions; and differences in reference points (white values used as anchors for other color properties). We will not need all of these details, but a summary of these color spaces is necessary, since I will be using many of them below.

It is useful to pause for a moment, however, and consider that color spaces are themselves descriptions, in the literary sense, of color—they can approximate their object, but only asymptotically. And yet so much of our modern world is made up of these approximations. Every screen we use—including the one I’m using to type this now, and the one you’re using to read it—is composed of millions of tiny picture elements—pixels—each with three lights: a red, a green, and a blue. The more specific colors, and the images, that we see on these screens are only mixtures of those elements.

The most common color spaces in use today include RGB, which stands for red, green, and blue; CMYK, or cyan, magenta, yellow, and white; HSL, or hue, saturation, and luminosity; and CIELAB, the newest and most accurate of these spaces. RGB is most common among light-producing devices like computer monitors, and generates colors additively, by mixing red, green, and blue light. These values are often expressed in hexidecimal, with the marker #, such that #ff0000, red, indicates the highest value for red (ff), along with the lowest value for green (00), and the lowest value for blue (00). CMYK is the most common for print media, on the other hand, since it describes colors subtractively, combining cyan, magenta, and yellow. HSL is a useful derivative of RGB, meanwhile, which allows for numeric manipulation of colors according to these values of hue, saturation, and luminosity.

The current standard colorspace, CIE \(L^* a^* b^*\), usually abbreviated CIELAB, is a product of a century’s long effort by the Commission Internationale de l’Eclairage [International Commission on Illumination], or CIE, an organization formed in 1913 to solve problems of chromaticity standardization, among others. A 1973 meeting of the CIE Colorimetry Committee, having evaluated a number of previously used color difference formulae, produced the first iteration of the LAB colorspace, intended to model human color perception. Here, \(L^*\) represents luminosity, \(a^*\) represents a spectrum of hues between green and magenta, and \(b^*\) represents hues between blue and yellow.

Relations between colors may then be calculated with respect to this coordinate system. The Euclidean distance between two colors in a LAB vector is therefore the square root of the differences of each of its components. The CIE calls this formula \(\Delta E\) (Robertson 167).

\[\Delta E = \sqrt{(\Delta L^*)^2 + (\Delta a^*)^2 + (\Delta b^*)^2}\]

Since CIELAB space best represents human perception of color, I’ll use it whereever possible, and calculate color distances using \(\Delta E\).I implement this function here, in the color categorization module of my color analyzer.

However, I have to translate frequently between LAB space and RGB space, since most of the color maps I’ve derived, are either scanned using digital photography, or, in the case of the XKCD map, produced using computer monitors.

Debates in Color Categories and Nomenclature

If we aim to quantify the occurences of certain color concepts, and not just the color words, then there must be a way to categorize visual experiences. For instance, if we encounter the expression light blue, we must be able to categorize this as a variety of blue, or else we will need to process and compare thousands of variables, instead of just a few. Yet the epistemological problems of the color/word interchange make this a difficult task. To begin with, since we are dealing with spectra, the boundaries of these categories are not well-defined. But the very existence of the categories themselves should not be assumed, either. While, to a painter or interior designer, the differences between ecru and eggshell may be crucial, these words may not be in the working vocabularies of some novelists. I say “working” here because they might be recognizable, and even familiar, to a writer, but they might not be the operative metaphors he or she chooses when describing a scene, or allowing a literary persona to describe it. So the color spectrum of a writer’s idiolect is always a subset of his or her dialect.

For instance, we might consider light blue to be a subcategory of blue, since the word blue is contained within it. However, is pink necessarily its own category, or is it simply a shade of red? And if so, is light pink a subcategory of pink or of red? We might categorize these colors differently if we were to use the hues rather than their written expressions.

We might look to other languages to see how these concepts are expressed, and learn about our own by comparison. Some languages lack a monolexemic term for pink, and others still have additional pink-like lexemes in other hue spectra. The Russian language, for instance, has the color-categories, or monolexemic color terms, синий [sínij], usually translated as “blue” or “dark blue,” and Голубо́й [golubój] which we might gloss as “light blue,” or “sky blue.” The image-based color mapping model, described below, predicts similar, but not identical colors for these English and Russian words, as well as their most common French translations:

Semantically and chromatically, these color categories are not synonymous. Just in the way that every translation requires some compromise, some reshaping, colors do not always cleanly map across languages. Some do: English blue and French bleu, as etymological kin, are not only morphologically closer than the English/Russian pair, but semantically, as well, and the model predicts this kinship.

The differences in color terminology between languages are important for us to bear in mind, even when the primary analysis below deals only with texts in English, because these differences are analogs for the gaps, and communications, between language and vision. Furthermore, most of the writers I’ll be discussing here speak more than one language: either from birth, as with Conrad and his native Polish, or through study, as with James Joyce, who was fluent at least five languages. And there have been some experiments in psychology that show semantic shifts in color categorization among speakers of more than one language (Ervin; Caskey-Sirmons and Hickerson; Athanasopoulos et al.).

More importantly, however, in order to categorize color words, we must first decide what our base color categories will be. This is no easy matter, and has long been the subject of debate. By comparing languages, linguists have often tried to ascertain what fundamental colors are, irrespective of their respective cultures.

One side of this debate calls into quesion the basis of fundamental colors, instead positing that color nomenclature, along with other phenomena, is in fact a cultural or linguistic construct. Probably the most well-known of these theories of linguistic relativism is that independently promoted, starting around the 1930s, by linguists Edward Sapir and Benjamin Whorf. Whorf’s 1940 summary of this view puts it succinctly: “the categories and types that we isolate from the world of phenomena we do not find there because they stare every observer in the face. On the contrary the world is presented in a kaleidoscopic flux of impressions which have to be organized in our minds. This means, largely, by the linguistic system in our minds” (Whorf 212).

On the other side of the debate, usually termed universalism, is an influential study of cross-linguistic color terminology, in a 1969 monograph of Brent Berlin and Paul Kay, Basic Color Terms: Their Universality and Evolution (Berlin and Kay 2). In particular, they name eleven categories: “white, black, red, green, yellow, blue, brown, purple, pink, orange, and grey,” and suggest that these categories develop in roughly that order—that all languages have words for white and black, that if they have a third, it is red, and so on.

Graphically, Berlin and Kay present this sequence as in the following diagram, where languages that have red must have both white and black, and so on. There is no order between yellow and green, but languages that develop a word for green would then develop a word for yellow, and vice-versa.

\[[\substack{white \\ black }] < [red] < [\substack{green \\ yellow}] < [blue] < [brown] < [\substack{purple \\ pink \\ orange \\ grey}]\]

Berlin and Kay see this sequence as a linguistic evolution in more than one sense—a dangerous term, in that it suggests a linear progression of simple to complex languages. The reasons they give for this are “increasing technological and cultural advancement” among the languages they compare. By way of explanation, they suppose that,

… to a group whose members have frequent occasion to contrast fine shades of leaf color and who possess no dyed fabrics, color-coded electrical wires, and so forth, it may not be worthwhile to rote-learn labels for gross perceptual discriminations such as green/blue, despite the psychophysical salience of such contrasts. (Berlin and Kay 16)

While the contrasts might have so physical basis, their linguistic categories do not map evenly to them, as Berlin and Kay themselves show. And as one might predicted, since 1969, their arguments of universal categories—and to a larger extent those of language evolution—have been either denounced as Anglocentric, or at least treated with a healthy skepticism. For instance, in 2006, Anna Wiezbicka argues that even the notion of color itself is not universal. Citing decades of research within the subfield of Natural Semantic Metalanguage, Wiezbicka argues that, “while many languages do not have a word for ‘colour,’ all languages have a word for seeing,’” and that “it makes more sense to ask about the universals of seeing rather than any putative ‘universals of colour’” (Wierzbicka 3).

I take no sides in this debate, but present it as evidence of the bond between language and perception. While there are few hard-line Whorfians remaining in linguistics, or universalists, both camps seem to agree that there are exceptions pulling at their theoretical sweaters, and colors are frequently the axis along which that pulling happens.

Is blood red?

To further complicate our conception of color/word translation, let’s return to the discussion of color word conventionality begun in the section on lemon-yellow above. As I have argued, although red is the conventional color of blood, blood itself is rarely red. This phenomenon presents itself in process of computational color categorization attempted here. While categorizing colors using CIELAB \(\Delta E\), which model human perception, I find that the category for the \(CM_X\) color word blood (#770001) gets categorized as brown, instead of red, as one might have predicted. Incidentally, blood red (#980002) is an entirely different color in the \(CM_X\), which is redder (i.e., contains a higher R value in its RGB representation) than blood. And dried blood (#4b0101) also exists, and is mapped to a darker red.

My initial feeling was that blood was miscategorized as a brown, and should instead be categorized as red. We all know blood is red–the term blood red itself proves it, right? But to look through images of blood, we may, in fact, discover that it is not red, but at best, a reddish brown. This is seemingly confirmed by the deep imaginer’s image-based imagined color (described below), which is #915b47. An image search at a stock photo provider like Unsplash or Pexels seems to confirm this, as well. However, crucially, the same searches for illustrations, rather than photos, depict blood as a bright red, instead of reddish brown—this seems to show that the linguistic-cognitive concept of blood is aligned with the concept of red, even though they aren’t visually equivalent. So when the OED editors, however meticulously they document the usages of blood-red, which date back to early Old English, gloss the term disappointingly literally as “red like blood; blood-coloured,” (OED, “blood-red”), [TODO: cite] they do not account for the discrepancy between the color of “blood-red” and the actual color of blood.

In British literature of this period, blood-red is often used to evoke other qualities of blood itself, although not necessarily its true color. In the hell-sermon that is the pivotal scene in Joyce’s A Portrait of the Artist as a Young Man, it is used to underscore the apocalyptic scene that Father Arnall is trying to describe: “the doomsday was at hand. The stars of heaven were falling upon the earth … The sun, … had become as sackcloth of hair. The moon was bloodred” [TODO: cite 99]. Lunar eclipses, in which the sun’s light on the moon is eclipsed, leaving only the earth’s light, make the moon appear dark red. These have long been described in English as a blood moon, but this is not just a color comparison: it is a metaphor which anthropomorphizes the moon in this state, comparing the moon’s face to one whose face has filled with blood, out of anger or another heightened emotional state. Father Arnall’s use of this metaphor, along with his simile for the sun, anthropomorphize heaven as a way to dramatize the wrath of God.

In Thomas Hardy’s Tess of the d’Urbervilles, Tess is described, in an early foreshadowing scene, as “not divining” that Alec d’Urberville, “one who stood fair to be the blood-red ray in the spectrum of her young life,” would come to be “the tragic mischief of her drama” [TODO: Broadview 73]. As in Joyce, “blood-red” allows for polysemy. First, it is “red … in the spectrum of her life”: red is the first, highest-frequency, and longest-wavelength band of a prismatic or spectrographic projection of Tess’s life, which implies that Alec will be for her among the first and most striking bands of her life. Spectroscopy—a kind of scientific “divining” of the material composition of matter, based on the spectral composition of its light—had come of age as a science in the 1870s and 80s, only a decade or two before Tess’s publication.

Second, “blood red” here implies a more literal red which comes from blood: a blushing which is seen in human faces, as well as, by extension of the metaphor, flowers, and fruit. This is the culmination of a chapter’s worth of red imagery, since Tess and Alec have just been picking strawberries and roses, and it is intertwined with imagery of Tess’s coming-of-age, or blossoming as the floral metaphor often has it.

When blood-red is understood as blushing, however, this is not the color of external, disembodied blood, which we have already established is more akin to brown, but refers to pinkish, blood-rich skin. In the Hungarian language, to choose one cross-cultural example, there are famously two words for red, vörös, derived from the word for blood, and piros, of similar etymology, but referring instead to, as linguist Anna Wierzbicka posits, “the color of blood inside a person’s body (visible sometimes in an open wound and in a person’s ‘red’ face)” (wiezbicka2006semantics?). In fact, Berlin and Kay note that many words for red are derived from blood (Berlin and Kay 38).

A red face, Wierzbica suggests, is not an attempt at accurately describing the color of someone’s face, but only that it has become more pink, i.e., taken on a more reddish hue than before. The red in question, then, is more of a reference to the concept of red, via blood and blood-red, than to the color phenomenon itself.

This red—again, not really the color red, but the concept—is the same red of rouge, the cosmetic used to emulate blushing, and whose name is derived from the French word for red. Rouge itself is often not red, but a somewhat reddish, pinkish, or purplish tint of another color. Max Beerbohm famously sings the praises of rouge, as a symbol of colorfulness and artifice, in an 1894 polemic in the short-lived by influential aestheticist journal bearing the name of another bright color: The Yellow Book. “The Pervasion of Rouge,” originally titled “A Defence of Cosmetics,” declares the end of the Victorian era, and thus “sancta simplicitas,” which we might interpret as a restricted color palette [TODO: cite]. Queen Victoria would not die, taking her eponymous era with her, for another seven years, but this declaration is an important herald of the “bright modernity” to come, as Blasszczyk and Spiekermann term it (Blaszczyk and Spiekermann).

P.A. Saccardo’s taxonomy does not place the color of blood with red at all, however, but with purple: he gives sanguineus as a Latin synonym of purpureus, along with the Greco-latin hæmatochrous, hæmatinus, and hæmatites (Saccardo 8). This is the traditional categorization of classical antiquity: the mapping appears in Homer, where in the Iliad, the earth is wet with purple blood. A. T. Murray’s English translation of Homer gives “thus mighty Aias charged them, and the earth grew wet with dark blood,” [αἵματι δὲ χθὼν δεύετο πορφυρέῳ] although πορφυρέῳ, which is translated as dark, is an etymological ancestor of purple [TODO: cite Perseus project here]. This categorization continues through Vergil, Ovid, and Horace. In fact, as Jacquiline Clarke points out, Horace plays with the traditional Homeric association of πορφύρεος with the sea and with death (πορφύρεος θάνατος, purple death or dark death, appears thrice in the Iliad), by juxtaposing the two in a purple blood-stained sea (Clarke 132). However, Liddell and Scott are quick to warn that “Homer seems not to have known the πορφύρα, [a purple fish, or purple dye] so that the word does not imply any definite colour.” [TODO: cite this purseus page].

To further complicate matters, Saccardo’s purpureus, while certainly on a spectrum that seems to range from red, to purple, and finally to brown, has a color of #8D0202, at least as it appears in the scanned edition from archive.org, however faded its original pigments may be. Some may rightly call this color red. So blood is not really red; it’s purple. But purple is red.

We may add blood to the long list of things called red which aren’t: the Red Sea (it’s blue), red wine and red grapes (they’re purple), red hair, red pandas, and Mars, the red planet (they’re orange). It comes as no surprise to report that red hair appears close to 900 times in \(C_{PG}\), but that orange hair, ginger hair, and copper hair are used only thrice each. Or that red wine appears 160 times, but purple wine only thrice. These are simply English-language conventions. But they prove that we must be especially careful while modeling our imagination of these terms. When we read red wine, do we imagine something red, or purple?

Similarly, when we read red hair, are we imagining red? The persistence of the villanous red-haired minor character trope in sensational literature of this period is evidence that the associations we’ve so far catalogued for red—blood, violence, ferocity, and so on—seep subconsciously into the characters’ depictions in fiction.Depictions of red-haired people in fiction could easily be the subject of another chapter, but are a little too far afield for this one. A concordance of \(C_{PG}\) for red-haired and similar terms, however, shows a multitude of unflattering accompanying personal descriptions.

We do not see those same stereotyped character attributes among more conscious and nuanced descriptions of hair color.

I want to reiterate here that these difficulties of textual color are not, as they may seem, merely background linguistic components of a literary art that is unconscious of them. Rather, they are fundamental to the process of literary description. Some writers are more explicit about these optical mechanics, and others are more implicit about them. But the modernist writers I choose to study the deepest in this chapter foreground color epistemologies in a way that, while it may not be a new literary device, is stronger, and brighter, and more variegated than before.

The wine-dark sea

Among the writers that deal most explicitly with color is James Joyce, an author I return to frequently. The first scene of Ulysses introduces a motif that recurs throughout the novel: the color of the sea. There, Buck Mulligan is gazing out onto the Irish sea from the crenellated parapets of Martello tower, in Sandycove, south of Dublin, and musing at once irreverently and reverently:

God! he said quietly. Isn’t the sea what Algy calls it: a great sweet mother? The snotgreen sea. The scrotumtightening sea. Epi oinopa ponton. Ah, Dedalus, the Greeks! I must teach you. You must read them in the original. Thalatta! Thalatta! She is our great sweet mother. Come and look. (Joyce, Ulysses 2)

How is the sea “snotgreen?” \(CM_X\) contains several colors for sea, as shown in table 4 below, as well as two mappings for snot. (Snot is not present in other color maps—unsurprisingly, since it would not very likely be a marketable name for a paint.)

And this list does not even include the many seafoam and seaweed variations. The variety of sea-like colors is an interesting problem, because seas themselves have a very wide range of colors among them, and even within any given sea. As suggested here in the name deep sea blue, the depth of the sea changes its apparent color. For comparison, the image-based color model predicts #98B8B3 for irish sea —a somewhat snot-green color.

Epi oinopa ponton, according to Don Gifford’s notes for Ulysses, is Homeric Greek for “upon the wine-dark sea,” a classic Homeric epithet that occurs throughout The Odyssey (Gifford and Seidman 15). It has long been a puzzle of Homeric scholarship as to why the sea is not blue, or green, but “wine-dark.” We should remember, however, that “dark” is an artifact of this translation convention, for in the Greek, which Mulligan advisedly does not gloss, “ἐπὶ οἴνοπα πόντον” might also be rendered “over the vinaceous sea” or “over the wine-like sea,” since οἴνοπα itself, despite clearly being used as a visual metaphor elsewhere in Homer, does not explicitly contain a signifier for dark, which would be closer to μέλας in Homeric Greek—in fact, elsewhere in Homer, wine itself is described as μέλας, although not here [Gladstone 472; TODO: cite LSJ].

One of the more well-known works of scholarship on this topic, however dated it may be considered now, is that put forth in William Gladstone’s 1858 Studies on Homer and the Homeric Age. Athough better known as the four-term prime minister of the United Kingdom, discontinuously from 1868 to 1894, Gladstone was a Homeric scholar of some distinction, and among the more interesting theses of this work is his catalog and interpretation of color words in the Homeric epics.

After a thorough concordance of visual terminology in Homer—which what one might half-jokingly call a 19th century digital humanities project—Gladstone concludes that Homer’s color expressions are relatively few. He lists as Homer’s only color words—excepting color metaphors—as λευκός (white), μέλας (black), ξανθός (yellow), έρυθρός (red), πορφύρεος (violet), κυάνεος (indigo), φοίνιξ (a phoenix, or Phoenician, purple or indigo), and πόλιος, (gray, grizzled) (Gladstone 459). His color metaphors, though, number thirteen, among which is οἴνοπα.

Gladstone notes that Homer applies οἴνοψ to only two objects, oxen and the sea. This puzzles him, however, since:

“there is no small difficulty in combining these two uses by reference to the idea of a common colour. The sea is blue, grey, or green. Oxen are black, bay, or brown. … It is remarkable that, among colours properly so called, Homer has none whatever, derived from the name of an object, that are light, unless it be in the case of the rose” (Gladstone 472).

οἴνοπα functions just as πορφύρεος does: as a visual descriptor of the sea, in the sense of “blood-red”: by comparing the sea to wine, it is not just the color that is compared, but other aspects, as well. We might imagine a tumultuous sea, for instance, which causes the ships upon it to sway as if drunken, as in Arthur Rimbaud’s poem “Le bateau ivre.” This same motion of the sea might also cause sailors on it to vomit as if they’d had too much wine.

The blood/wine/sea metaphoric trinity was not lost on Joyce, either: in the “Proteus” episode of Ulysses, we see Stephen daydream the following, looking again out at the Irish sea:

A tide westering, moondrawn, in her wake. Tides, myriadislanded, within her, blood not mine, oinopa ponton, a winedark sea. Behold the handmaid of the moon. In sleep the wet sign calls her hour, bids her rise. Bridebed, childbed, bed of death, ghostcandled. Omnis caro ad te veniet. He comes, pale vampire, through storm his eyes, his bat sails bloodying the sea, mouth to her mouth’s kiss.

Here, Stephen’s poetically free-associating imagination conjures a nighttime sea as “the handmaid of the moon,” because it is “pulled” by it in its tides. He extends this feminine analogy, via the conventional euphemism for menstruation, to a series of blood-soaked bedsheets, with their analogue in the bloodied sea, and a recollection of a sexual episode with a prostitute that Stephen will remember more fully later. A common connection in this stream—or sea?—of consciousness is the purple color.

What’s important here is to recognize that this purple is closer to how it appears than how it is categorized. Again, conventional associations have it that the sea is blue, and that blood is red, and that red wine is of course red, but to read Homeric descriptions of the sea and blood and wine as purple, we are more reminded of the perceptual phemomenon than the linguistic category. And that, I argue, is one of the primary projects of the modernist movement.

\(C_{PG}\): Project Gutenberg

\(C_{BL}\): British Library

\(C_{PGA}\): Project Gutenberg Australia

\(C_{NG}\): Google Books Fiction (Ngrams)

Corpus-DB and Metadata Augmentation